I have an Amazon EC2 deployment of mongodb (3.4). Things are going well, but the DB size is growing quickly. I am about to shard a large collection in order to begin horizontal scaling (total space needed is 2TB).

So either (ignoring config servers, etc):

1. Each shard is a replica set with 3 data nodes

total cost = $1,130 / mo

6x m4.large ($85 ea) + 4x 1TB SSDs ($116 ea) + 2x 1TB magnetic (to save some $$) ($78 ea)

2. Each shard is a replica set with 2 data nodes, each also using a shared arbiter

total cost = $809 / mo

4x m4.large ($85 ea) + 4x 1TB SSDs ($116 ea) + 1x arbiter (cheapest machine is $5)

Diff is $321/mo

I get the feeling that option 2, in Amazon's hosted environment, using exclusively SSDs, should be quite durable. As far as I can see, the only problem with Option 2 is that if a primary node dies and I failover to the secondary, for that period there is no backup. But I can't actually evaluate the severity of this scenario.

Could Option 2 be mitigated by attaching a spinning disk to each data node, that will act as a backup volume?

If anyone could provide some advice from experience it would be super helpful, but any advice is appreciated.



2 Answers 2


Option 2 has no problems in theory; however the more limited redundancy may be a problem.

  1. About the arbiters:

Yes, you can run multiple arbiters on a single lightweight machine. An arbiter is a very lightweight process, doing little but voting, so you can run multiple arbiters on a single box. Each one must be a separate mongod process.

  1. About the redundancy:

If you have a 2-data-node replica set, and one of the data nodes fails, then the other one can function as primary (with the vote of the arbiter), so you have resilience there, but you no longer have redundancy.

Whether that is a severe problem or not depends on your ops, how quickly you can fix or replace the failed node. That depends on your monitoring capabilities, your ops staff competence and availability, etc; that's a decision you need to evaluate.


With option 1, why you want to have four nodes in replica set? Three is enough. With four nodes, you should have arbiter as fifth, because vote count must be odd. With those three nodes, every node should be in the different data center, so you have DR and HA.

Minimum to have DR and HA, two data nodes and one arbiter, all in different data centers.

  • 2 shards, 4 nodes (TOTAL), so 1 shard + 2 replicas, each
    – mils
    Commented Nov 3, 2017 at 6:29
  • So, if your replica set is only 2 nodes, you need arbiter.. Without it, your (shard) replica set don't have primary if one node is down (one is not majority when votes are cast). Arbiter don't need (really) any disk space, so it can be very little virtual machine. So, still three DC setup is best. Two DC for data nodes and one for arbiter. Two AWS and one some cheap VPS provider what have $2/month XS-size VPS available.
    – JJussi
    Commented Nov 3, 2017 at 10:20
  • If a shard has two replicas, is this not a replica set containing three nodes?
    – mils
    Commented Nov 4, 2017 at 2:51
  • If shard have two nodes, it is two node replicaset (primary + secondary). If shard have three nodes, it is three node replicaset (primary + 2xsecondary). With two node replica set, to reach majority, you need two votes and only way to archive that is to have arbiter what's only purpose is vote, not save data or serve clients.
    – JJussi
    Commented Nov 4, 2017 at 9:24
  • Good, that means in Option 1 I have three nodes in a replica set, not two.
    – mils
    Commented Nov 5, 2017 at 22:48

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.